Multiple Working Conditions Soft Sensor Method Based On Geodesic Flow Kernel

Process industry is a pillar industry that occupies an important strategic position in the national economy.Typical process industries mainly include raw materials industries such as petroleum,chemical,and mineral processing.Due to the complex operation mechanism of the production line of the production industry,the system is limited by factors such as non-linearity,strong coupling,and large time lag.It makes it difficult to detect relevant key parameters in real time using hardware sensors,and cannot effectively monitor the operating status of production equipment,which seriously affects the smooth operation of the production process and product quality.Soft measurement technology is an effective method to solve the abovementioned problems.It can realize the real-time prediction of difficult-to-measure leading variables by constructing a mathematical model between easy-to-measure auxiliary variables and difficult-to-measure dominant variables in industrial production.In practice,changes in production tasks,equipment reorganization,changes in raw materials and on-site environment make the system reflect the characteristics of multiple operating conditions.The changes in operating conditions make the distribution of real-time data and historical modeling data not obey the same distribution,causing the performance of the original soft sensor model to decline.When most of the unknown working condition data is difficult to establish an accurate soft measurement model because of the inability to obtain labels,how to mine the effective information of the unknown working condition from the known working condition data and assist the modeling of the unknown working condition becomes an urgent problem to be solved.Based on the above analysis,the unsupervised transfer learning framework based on manifold is introduced into the field of soft measurement,and the geodesic flow kernel method is used as the basis of this study to solve the soft measurement of key parameters of unknown conditions in the process industry with multi-condition characteristics the main research work is as follows:(1)To deal with the problem of complex multi-working conditions in the process industry,the traditional soft sensor model is inaccurate and the new working conditions cannot be modeled without any labeled data.First extract the common mode information of different working conditions,then from the perspective of unsupervised transfer learning,build a multi-operation soft sensor model based on the extraction of common mode information on the basis of the geodesic flow kernel algorithm framework to realize the unknown Accurate prediction of key parameters in operating conditions.(2)In view of the dynamic characteristics existing in process industry data,the timing relationship of variables is taken into account,and the PCA augmentation matrix is introduced to solve the problem of data dynamic characteristics extraction.Aiming at the problem that the single domain migration method is difficult to use multiple historical working condition data information,a multi-source domain migration soft measurement strategy is introduced to build multiple soft measurement sub-models based on multiple historical known working conditions and unmolded working condition data,and integrate get the predicted value.At the same time,based on the ISOMAP method,the geodesic distance is used instead of the Euclidean distance to better extract the common mode information of different working conditions and improve the accuracy of the soft sensor model.(3)Aiming at the problem that the soft measurement model needs to be updated and corrected in a continuous industrial process,a multi-operation online soft measurement modeling method is studied.Combining real-time learning strategies with transfer learning methods,on the one hand,avoids the problem of poor model accuracy due to differences in data distribution between real-time learning conditions,and on the other hand,makes transfer learning soft-sensing methods self-adjusting.At the same time,the integrated learning strategy is introduced to increase the reliability of the algorithm.The experimental results and analysis verify the effectiveness and practicability of the proposed algorithm.